The invention provides a load characteristic self-learning method suitable for non-intrusive power monitoring. The method comprises the steps of obtaining a load data sequence after a load event occurs as an input sample; randomly adding noise into the sample; determining the number of input layer neurons of the noise reduction auto-encoder according to the length of the data sample sequence, andgenerating an input layer and an output layer; determining the number of neurons of the hidden layer of the auto-encoder and generating a hidden layer; setting a training error limit of the noise reduction auto-encoder; initializing inter-layer mapping parameters of the noise reduction auto-encoder; calculating a reconstruction error of the sequence for the input sequence according to the mappingparameter; and if the reconstruction error is smaller than the training error limit, extracting a hidden layer node value as an abstract feature of the load event, and if the reconstruction error is larger than the training error limit, updating mapping parameters between the input layer and the hidden layer and between the hidden layer and the output layer by utilizing a gradient descent algorithm. According to the method, the compressed sensing of the data sequence is realized, so that the learning of abstract features is realized, and the load event data curve is globally interpreted.